Merge branch 'bpf-cpumap-enable-gro-for-xdp_pass-frames'

Alexander Lobakin says:

====================
bpf: cpumap: enable GRO for XDP_PASS frames

Several months ago, I had been looking through my old XDP hints tree[0]
to check whether some patches not directly related to hints can be sent
standalone. Roughly at the same time, Daniel appeared and asked[1] about
GRO for cpumap from that tree.

Currently, cpumap uses its own kthread which processes cpumap-redirected
frames by batches of 8, without any weighting (but with rescheduling
points). The resulting skbs get passed to the stack via
netif_receive_skb_list(), which means no GRO happens.
Even though we can't currently pass checksum status from the drivers,
in many cases GRO performs better than the listified Rx without the
aggregation, confirmed by tests.

In order to enable GRO in cpumap, we need to do the following:

* patches 1-2: decouple the GRO struct from the NAPI struct and allow
  using it out of a NAPI entity within the kernel core code;
* patch 3: switch cpumap from netif_receive_skb_list() to
  gro_receive_skb().

Additional improvements:

* patch 4: optimize XDP_PASS in cpumap by using arrays instead of linked
  lists;
* patch 5-6: introduce and use function do get skbs from the NAPI percpu
  caches by bulks, not one at a time;
* patch 7-8: use that function in veth as well and remove the one that
  was now superseded by it.

My trafficgen UDP GRO tests, small frame sizes:

                GRO off    GRO on
baseline        2.7        N/A       Mpps
patch 3         2.3        4         Mpps
patch 8         2.4        4.7       Mpps

1...3 diff      -17        +48       %
1...8 diff      -11        +74       %

Daniel reported from +14%[2] to +18%[3] of throughput in neper's TCP RR
tests. On my system however, the same test gave me up to +100%.

Note that there's a series from Lorenzo[4] which achieves the same, but
in a different way. During the discussions, the approach using a
standalone GRO instance was preferred over the threaded NAPI.

[0] https://github.com/alobakin/linux/tree/xdp_hints
[1] https://lore.kernel.org/bpf/cadda351-6e93-4568-ba26-21a760bf9a57@app.fastmail.com
[2] https://lore.kernel.org/bpf/merfatcdvwpx2lj4j2pahhwp4vihstpidws3jwljwazhh76xkd@t5vsh4gvk4mh
[3] https://lore.kernel.org/bpf/yzda66wro5twmzpmjoxvy4si5zvkehlmgtpi6brheek3sj73tj@o7kd6nurr3o6
[4] https://lore.kernel.org/bpf/20241130-cpumap-gro-v1-0-c1180b1b5758@kernel.org
====================

Link: https://patch.msgid.link/20250225171751.2268401-1-aleksander.lobakin@intel.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>

4 files changed, 152 insertions, 102 deletions

diff --git a/net/core/dev.c b/net/core/dev.c
index b0ab0169f507..2dd9e6f490e6 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -6484,7 +6484,7 @@ bool napi_complete_done(struct napi_struct *n, int work_done)
 		return false;
 
 	if (work_done) {
-		if (n->gro_bitmask)
+		if (n->gro.bitmask)
 			timeout = napi_get_gro_flush_timeout(n);
 		n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n);
 	}
@@ -6494,15 +6494,14 @@ bool napi_complete_done(struct napi_struct *n, int work_done)
 		if (timeout)
 			ret = false;
 	}
-	if (n->gro_bitmask) {
-		/* When the NAPI instance uses a timeout and keeps postponing
-		 * it, we need to bound somehow the time packets are kept in
-		 * the GRO layer
-		 */
-		napi_gro_flush(n, !!timeout);
-	}
 
-	gro_normal_list(n);
+	/*
+	 * When the NAPI instance uses a timeout and keeps postponing
+	 * it, we need to bound somehow the time packets are kept in
+	 * the GRO layer.
+	 */
+	gro_flush(&n->gro, !!timeout);
+	gro_normal_list(&n->gro);
 
 	if (unlikely(!list_empty(&n->poll_list))) {
 		/* If n->poll_list is not empty, we need to mask irqs */
@@ -6566,19 +6565,15 @@ static void skb_defer_free_flush(struct softnet_data *sd)
 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule)
 {
 	if (!skip_schedule) {
-		gro_normal_list(napi);
+		gro_normal_list(&napi->gro);
 		__napi_schedule(napi);
 		return;
 	}
 
-	if (napi->gro_bitmask) {
-		/* flush too old packets
-		 * If HZ < 1000, flush all packets.
-		 */
-		napi_gro_flush(napi, HZ >= 1000);
-	}
+	/* Flush too old packets. If HZ < 1000, flush all packets */
+	gro_flush(&napi->gro, HZ >= 1000);
+	gro_normal_list(&napi->gro);
 
-	gro_normal_list(napi);
 	clear_bit(NAPI_STATE_SCHED, &napi->state);
 }
 
@@ -6685,7 +6680,7 @@ restart:
 		}
 		work = napi_poll(napi, budget);
 		trace_napi_poll(napi, work, budget);
-		gro_normal_list(napi);
+		gro_normal_list(&napi->gro);
 count:
 		if (work > 0)
 			__NET_ADD_STATS(dev_net(napi->dev),
@@ -6785,6 +6780,8 @@ void napi_resume_irqs(unsigned int napi_id)
 static void __napi_hash_add_with_id(struct napi_struct *napi,
 				    unsigned int napi_id)
 {
+	napi->gro.cached_napi_id = napi_id;
+
 	WRITE_ONCE(napi->napi_id, napi_id);
 	hlist_add_head_rcu(&napi->napi_hash_node,
 			   &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
@@ -6853,17 +6850,6 @@ static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
 	return HRTIMER_NORESTART;
 }
 
-static void init_gro_hash(struct napi_struct *napi)
-{
-	int i;
-
-	for (i = 0; i < GRO_HASH_BUCKETS; i++) {
-		INIT_LIST_HEAD(&napi->gro_hash[i].list);
-		napi->gro_hash[i].count = 0;
-	}
-	napi->gro_bitmask = 0;
-}
-
 int dev_set_threaded(struct net_device *dev, bool threaded)
 {
 	struct napi_struct *napi;
@@ -7191,10 +7177,8 @@ void netif_napi_add_weight_locked(struct net_device *dev,
 	INIT_HLIST_NODE(&napi->napi_hash_node);
 	hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
 	napi->timer.function = napi_watchdog;
-	init_gro_hash(napi);
+	gro_init(&napi->gro);
 	napi->skb = NULL;
-	INIT_LIST_HEAD(&napi->rx_list);
-	napi->rx_count = 0;
 	napi->poll = poll;
 	if (weight > NAPI_POLL_WEIGHT)
 		netdev_err_once(dev, "%s() called with weight %d\n", __func__,
@@ -7308,19 +7292,6 @@ void napi_enable(struct napi_struct *n)
 }
 EXPORT_SYMBOL(napi_enable);
 
-static void flush_gro_hash(struct napi_struct *napi)
-{
-	int i;
-
-	for (i = 0; i < GRO_HASH_BUCKETS; i++) {
-		struct sk_buff *skb, *n;
-
-		list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list)
-			kfree_skb(skb);
-		napi->gro_hash[i].count = 0;
-	}
-}
-
 /* Must be called in process context */
 void __netif_napi_del_locked(struct napi_struct *napi)
 {
@@ -7343,8 +7314,7 @@ void __netif_napi_del_locked(struct napi_struct *napi)
 	list_del_rcu(&napi->dev_list);
 	napi_free_frags(napi);
 
-	flush_gro_hash(napi);
-	napi->gro_bitmask = 0;
+	gro_cleanup(&napi->gro);
 
 	if (napi->thread) {
 		kthread_stop(napi->thread);
@@ -7403,14 +7373,9 @@ static int __napi_poll(struct napi_struct *n, bool *repoll)
 		return work;
 	}
 
-	if (n->gro_bitmask) {
-		/* flush too old packets
-		 * If HZ < 1000, flush all packets.
-		 */
-		napi_gro_flush(n, HZ >= 1000);
-	}
-
-	gro_normal_list(n);
+	/* Flush too old packets. If HZ < 1000, flush all packets */
+	gro_flush(&n->gro, HZ >= 1000);
+	gro_normal_list(&n->gro);
 
 	/* Some drivers may have called napi_schedule
 	 * prior to exhausting their budget.
@@ -12439,7 +12404,7 @@ static struct hlist_head * __net_init netdev_create_hash(void)
 static int __net_init netdev_init(struct net *net)
 {
 	BUILD_BUG_ON(GRO_HASH_BUCKETS >
-		     8 * sizeof_field(struct napi_struct, gro_bitmask));
+		     BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask));
 
 	INIT_LIST_HEAD(&net->dev_base_head);
 
@@ -12804,7 +12769,7 @@ static int __init net_dev_init(void)
 		INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd);
 		spin_lock_init(&sd->defer_lock);
 
-		init_gro_hash(&sd->backlog);
+		gro_init(&sd->backlog.gro);
 		sd->backlog.poll = process_backlog;
 		sd->backlog.weight = weight_p;
 		INIT_LIST_HEAD(&sd->backlog.poll_list);
diff --git a/net/core/gro.c b/net/core/gro.c
index 78b320b63174..19bd4cdaee3a 100644
--- a/net/core/gro.c
+++ b/net/core/gro.c
@@ -250,8 +250,7 @@ int skb_gro_receive_list(struct sk_buff *p, struct sk_buff *skb)
 	return 0;
 }
 
-
-static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
+static void gro_complete(struct gro_node *gro, struct sk_buff *skb)
 {
 	struct list_head *head = &net_hotdata.offload_base;
 	struct packet_offload *ptype;
@@ -284,43 +283,43 @@ static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
 	}
 
 out:
-	gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
+	gro_normal_one(gro, skb, NAPI_GRO_CB(skb)->count);
 }
 
-static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
-				   bool flush_old)
+static void __gro_flush_chain(struct gro_node *gro, u32 index, bool flush_old)
 {
-	struct list_head *head = &napi->gro_hash[index].list;
+	struct list_head *head = &gro->hash[index].list;
 	struct sk_buff *skb, *p;
 
 	list_for_each_entry_safe_reverse(skb, p, head, list) {
 		if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
 			return;
 		skb_list_del_init(skb);
-		napi_gro_complete(napi, skb);
-		napi->gro_hash[index].count--;
+		gro_complete(gro, skb);
+		gro->hash[index].count--;
 	}
 
-	if (!napi->gro_hash[index].count)
-		__clear_bit(index, &napi->gro_bitmask);
+	if (!gro->hash[index].count)
+		__clear_bit(index, &gro->bitmask);
 }
 
-/* napi->gro_hash[].list contains packets ordered by age.
+/*
+ * gro->hash[].list contains packets ordered by age.
  * youngest packets at the head of it.
  * Complete skbs in reverse order to reduce latencies.
  */
-void napi_gro_flush(struct napi_struct *napi, bool flush_old)
+void __gro_flush(struct gro_node *gro, bool flush_old)
 {
-	unsigned long bitmask = napi->gro_bitmask;
+	unsigned long bitmask = gro->bitmask;
 	unsigned int i, base = ~0U;
 
 	while ((i = ffs(bitmask)) != 0) {
 		bitmask >>= i;
 		base += i;
-		__napi_gro_flush_chain(napi, base, flush_old);
+		__gro_flush_chain(gro, base, flush_old);
 	}
 }
-EXPORT_SYMBOL(napi_gro_flush);
+EXPORT_SYMBOL(__gro_flush);
 
 static unsigned long gro_list_prepare_tc_ext(const struct sk_buff *skb,
 					     const struct sk_buff *p,
@@ -439,7 +438,7 @@ static void gro_try_pull_from_frag0(struct sk_buff *skb)
 		gro_pull_from_frag0(skb, grow);
 }
 
-static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
+static void gro_flush_oldest(struct gro_node *gro, struct list_head *head)
 {
 	struct sk_buff *oldest;
 
@@ -455,14 +454,15 @@ static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
 	 * SKB to the chain.
 	 */
 	skb_list_del_init(oldest);
-	napi_gro_complete(napi, oldest);
+	gro_complete(gro, oldest);
 }
 
-static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
+static enum gro_result dev_gro_receive(struct gro_node *gro,
+				       struct sk_buff *skb)
 {
 	u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
-	struct gro_list *gro_list = &napi->gro_hash[bucket];
 	struct list_head *head = &net_hotdata.offload_base;
+	struct gro_list *gro_list = &gro->hash[bucket];
 	struct packet_offload *ptype;
 	__be16 type = skb->protocol;
 	struct sk_buff *pp = NULL;
@@ -526,7 +526,7 @@ found_ptype:
 
 	if (pp) {
 		skb_list_del_init(pp);
-		napi_gro_complete(napi, pp);
+		gro_complete(gro, pp);
 		gro_list->count--;
 	}
 
@@ -537,7 +537,7 @@ found_ptype:
 		goto normal;
 
 	if (unlikely(gro_list->count >= MAX_GRO_SKBS))
-		gro_flush_oldest(napi, &gro_list->list);
+		gro_flush_oldest(gro, &gro_list->list);
 	else
 		gro_list->count++;
 
@@ -551,10 +551,10 @@ found_ptype:
 	ret = GRO_HELD;
 ok:
 	if (gro_list->count) {
-		if (!test_bit(bucket, &napi->gro_bitmask))
-			__set_bit(bucket, &napi->gro_bitmask);
-	} else if (test_bit(bucket, &napi->gro_bitmask)) {
-		__clear_bit(bucket, &napi->gro_bitmask);
+		if (!test_bit(bucket, &gro->bitmask))
+			__set_bit(bucket, &gro->bitmask);
+	} else if (test_bit(bucket, &gro->bitmask)) {
+		__clear_bit(bucket, &gro->bitmask);
 	}
 
 	return ret;
@@ -593,13 +593,12 @@ struct packet_offload *gro_find_complete_by_type(__be16 type)
 }
 EXPORT_SYMBOL(gro_find_complete_by_type);
 
-static gro_result_t napi_skb_finish(struct napi_struct *napi,
-				    struct sk_buff *skb,
-				    gro_result_t ret)
+static gro_result_t gro_skb_finish(struct gro_node *gro, struct sk_buff *skb,
+				   gro_result_t ret)
 {
 	switch (ret) {
 	case GRO_NORMAL:
-		gro_normal_one(napi, skb, 1);
+		gro_normal_one(gro, skb, 1);
 		break;
 
 	case GRO_MERGED_FREE:
@@ -620,21 +619,21 @@ static gro_result_t napi_skb_finish(struct napi_struct *napi,
 	return ret;
 }
 
-gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
+gro_result_t gro_receive_skb(struct gro_node *gro, struct sk_buff *skb)
 {
 	gro_result_t ret;
 
-	skb_mark_napi_id(skb, napi);
+	__skb_mark_napi_id(skb, gro);
 	trace_napi_gro_receive_entry(skb);
 
 	skb_gro_reset_offset(skb, 0);
 
-	ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
+	ret = gro_skb_finish(gro, skb, dev_gro_receive(gro, skb));
 	trace_napi_gro_receive_exit(ret);
 
 	return ret;
 }
-EXPORT_SYMBOL(napi_gro_receive);
+EXPORT_SYMBOL(gro_receive_skb);
 
 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
 {
@@ -690,7 +689,7 @@ static gro_result_t napi_frags_finish(struct napi_struct *napi,
 		__skb_push(skb, ETH_HLEN);
 		skb->protocol = eth_type_trans(skb, skb->dev);
 		if (ret == GRO_NORMAL)
-			gro_normal_one(napi, skb, 1);
+			gro_normal_one(&napi->gro, skb, 1);
 		break;
 
 	case GRO_MERGED_FREE:
@@ -759,7 +758,7 @@ gro_result_t napi_gro_frags(struct napi_struct *napi)
 
 	trace_napi_gro_frags_entry(skb);
 
-	ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
+	ret = napi_frags_finish(napi, skb, dev_gro_receive(&napi->gro, skb));
 	trace_napi_gro_frags_exit(ret);
 
 	return ret;
@@ -791,3 +790,37 @@ __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
 	return sum;
 }
 EXPORT_SYMBOL(__skb_gro_checksum_complete);
+
+void gro_init(struct gro_node *gro)
+{
+	for (u32 i = 0; i < GRO_HASH_BUCKETS; i++) {
+		INIT_LIST_HEAD(&gro->hash[i].list);
+		gro->hash[i].count = 0;
+	}
+
+	gro->bitmask = 0;
+	gro->cached_napi_id = 0;
+
+	INIT_LIST_HEAD(&gro->rx_list);
+	gro->rx_count = 0;
+}
+
+void gro_cleanup(struct gro_node *gro)
+{
+	struct sk_buff *skb, *n;
+
+	for (u32 i = 0; i < GRO_HASH_BUCKETS; i++) {
+		list_for_each_entry_safe(skb, n, &gro->hash[i].list, list)
+			kfree_skb(skb);
+
+		gro->hash[i].count = 0;
+	}
+
+	gro->bitmask = 0;
+	gro->cached_napi_id = 0;
+
+	list_for_each_entry_safe(skb, n, &gro->rx_list, list)
+		kfree_skb(skb);
+
+	gro->rx_count = 0;
+}
diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index 5b241c9e6f38..f12815f9c83d 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -295,6 +295,68 @@ static struct sk_buff *napi_skb_cache_get(void)
 	return skb;
 }
 
+/**
+ * napi_skb_cache_get_bulk - obtain a number of zeroed skb heads from the cache
+ * @skbs: pointer to an at least @n-sized array to fill with skb pointers
+ * @n: number of entries to provide
+ *
+ * Tries to obtain @n &sk_buff entries from the NAPI percpu cache and writes
+ * the pointers into the provided array @skbs. If there are less entries
+ * available, tries to replenish the cache and bulk-allocates the diff from
+ * the MM layer if needed.
+ * The heads are being zeroed with either memset() or %__GFP_ZERO, so they are
+ * ready for {,__}build_skb_around() and don't have any data buffers attached.
+ * Must be called *only* from the BH context.
+ *
+ * Return: number of successfully allocated skbs (@n if no actual allocation
+ *	   needed or kmem_cache_alloc_bulk() didn't fail).
+ */
+u32 napi_skb_cache_get_bulk(void **skbs, u32 n)
+{
+	struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
+	u32 bulk, total = n;
+
+	local_lock_nested_bh(&napi_alloc_cache.bh_lock);
+
+	if (nc->skb_count >= n)
+		goto get;
+
+	/* No enough cached skbs. Try refilling the cache first */
+	bulk = min(NAPI_SKB_CACHE_SIZE - nc->skb_count, NAPI_SKB_CACHE_BULK);
+	nc->skb_count += kmem_cache_alloc_bulk(net_hotdata.skbuff_cache,
+					       GFP_ATOMIC | __GFP_NOWARN, bulk,
+					       &nc->skb_cache[nc->skb_count]);
+	if (likely(nc->skb_count >= n))
+		goto get;
+
+	/* Still not enough. Bulk-allocate the missing part directly, zeroed */
+	n -= kmem_cache_alloc_bulk(net_hotdata.skbuff_cache,
+				   GFP_ATOMIC | __GFP_ZERO | __GFP_NOWARN,
+				   n - nc->skb_count, &skbs[nc->skb_count]);
+	if (likely(nc->skb_count >= n))
+		goto get;
+
+	/* kmem_cache didn't allocate the number we need, limit the output */
+	total -= n - nc->skb_count;
+	n = nc->skb_count;
+
+get:
+	for (u32 base = nc->skb_count - n, i = 0; i < n; i++) {
+		u32 cache_size = kmem_cache_size(net_hotdata.skbuff_cache);
+
+		skbs[i] = nc->skb_cache[base + i];
+
+		kasan_mempool_unpoison_object(skbs[i], cache_size);
+		memset(skbs[i], 0, offsetof(struct sk_buff, tail));
+	}
+
+	nc->skb_count -= n;
+	local_unlock_nested_bh(&napi_alloc_cache.bh_lock);
+
+	return total;
+}
+EXPORT_SYMBOL_GPL(napi_skb_cache_get_bulk);
+
 static inline void __finalize_skb_around(struct sk_buff *skb, void *data,
 					 unsigned int size)
 {
diff --git a/net/core/xdp.c b/net/core/xdp.c
index 2c6ab6fb452f..f86eedad586a 100644
--- a/net/core/xdp.c
+++ b/net/core/xdp.c
@@ -618,16 +618,6 @@ void xdp_warn(const char *msg, const char *func, const int line)
 };
 EXPORT_SYMBOL_GPL(xdp_warn);
 
-int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
-{
-	n_skb = kmem_cache_alloc_bulk(net_hotdata.skbuff_cache, gfp, n_skb, skbs);
-	if (unlikely(!n_skb))
-		return -ENOMEM;
-
-	return 0;
-}
-EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk);
-
 /**
  * xdp_build_skb_from_buff - create an skb from &xdp_buff
  * @xdp: &xdp_buff to convert to an skb